Why are advanced language models incapable of making money on the stock market? As research authored by Huangfei Jiang, Anning Hu, Kai Sheng, Yuchen Cai, Rui Ge, Xiangyu Kong, Jianghao Lin, Shuzi Chen, Jiyi Weng, and Junchi Yan demonstrates, the capacity to generate polished analytical text no longer guarantees the accuracy of financial forecasts and investment returns. Professional equity research requires rigorous decision-time evidence evaluation, yet even top systems falter when moving from static knowledge retrieval to autonomous capital management.
To measure this structural limitation, researchers introduced Investor's Last Exam (ILE), an equity-research benchmark led by experienced buy-side investors and fund managers. As detailed in the paper published on 28 Oct 2026 under the identifier arXiv:2610.38342v1, the benchmark evaluates agents across 86 assignments, jointly assessing research quality, forecast accuracy, and simulated investment performance. The evaluation framework relies on a versioned financial database spanning 3.43 million document versions and nine source categories, supplemented by structured market data with validated timestamps governing retrieval and computation, establishing a strictly frozen information set and professional evaluation criteria to test agents without looking into the future.
The Architecture of Financial Failure
When deploying AI models for market analysis, executive teams often assume that high general scores translate directly to trading success. Yet flagship systems reveal a stark divergence between analytical fluency and predictive performance. GPT-6 Astra and Claude Opus 5.5 lead overall at 68.1 and 66.9 out of 100 in the evaluation, establishing strong baselines in general research tasks. However, even Astra scores only 52.5 on Forecast & Model, highlighting a severe drop-off when transitioning from qualitative summaries to quantitative projection.
GPT-6 Astra and Claude Opus 5.5 lead overall at 68.1 and 66.9 out of 100 in the evaluation, yet even Astra scores only 52.5 on Forecast & Model.
This performance gap stems from the complexity of the decision-making chain. Financial agents must analyze operations, competitive positions, and financial conditions while navigating strict dependencies like loan covenants and drawable credit versus unused credit. If an agent fails to trace these variables through financial dependency graphs, its simulation falls apart when facing subsequent disclosures revealed after the decision-time cutoff.
Moving Past Static Benchmarks
Some might argue that financial forecasting improves rapidly with scaling laws and larger datasets, making current benchmark failures temporary. How does the market respond when high analytical scores mask poor forecasting reliability?
As the data shows, high scores in answer quality and task performance do not prevent failure in outcome accuracy. The project page hosted at https://prophet-lab.cn/benchmarks/investor-last-exam-v1 and the accompanying code available at https://github.com/Prophetlab/Investor-Last-Exam provide the diagnostic foundation needed to locate where analytical advantages fail. Researchers and developers can examine the dataset available at https://huggingface.co/datasets/HenryJ12/InvestorLastExam to audit the exact evidence-assumption-judgment chain.
The shift from isolated capabilities to traceable evaluation changes how organizations assess autonomous systems. When capital allocation decisions depend on rigorous calculations and reconciliation, models must prove their predictive validity rather than merely reciting historical filings. As financial markets tighten operational budgets, the inability to translate research capability into reliable forecasts forces a reevaluation of automated investment strategies, leaving the original question of market profitability unresolved for autonomous systems.
What doesn’t add up
Can an enterprise safely hand over capital allocation to an LLM simply because it drafts fluent equity research? The reality proves otherwise: companies pay for analytical rhetoric while deploying agents that remain blind to hard structural links like credit covenants and uncommitted borrowing limits. When even a flagship like GPT-6 Astra sinks to fifty-two points in quantitative forecasting, the cost of confusing verbal literacy with financial predictive accuracy becomes painfully tangible. The frontier models compose immaculate investment memos, yet they cannot make money on their own advice.